Friday July 25, 2025

OpenAI prepares to launch GPT-5 in August, researchers find that Transformers can achieve similar performance without normalization layers, and Superglue enables users to integrate and orchestrate APIs using natural language.

News

Two narratives about AI

There are two conflicting narratives about the impact of AI on the programming industry, with some claiming it will revolutionize and automate coding, leading to significant job losses, while others argue that AI tools are not yet capable of replacing human programmers and may even hinder their productivity. Ultimately, the true effects of AI on the industry are still unknown, and it's best to approach the topic with a critical and nuanced perspective, tuning out extreme rhetoric and focusing on tangible changes and trusted sources.

OpenAI prepares to launch GPT-5 in August

OpenAI is preparing to launch its next-generation GPT-5 model in early August, with CEO Sam Altman recently teasing its capabilities and sources indicating that it will include mini and nano versions available through the company's API. The launch was initially expected in late May but was delayed for additional testing, and OpenAI has not officially commented on the new release date.

Train a 70b language model at home (2024)

Answer.AI has released a fully open source system that enables the efficient training of a 70 billion parameter large language model on a regular desktop computer with two or more standard gaming GPUs. This system, which combines FSDP and QLoRA, aims to make large model training more accessible and affordable, allowing individuals and small labs to create their own personalized models without relying on expensive data center hardware.

Why Are We Pretending AI Is Going to Take All the Jobs?

Executives and experts, including Jim Cramer and CEOs of major companies like Amazon, JP Morgan, and Ford, are warning of a "jobs apocalypse" due to the emergence of generative artificial intelligence, which they claim will displace millions of jobs, particularly in white-collar industries. However, the author argues that this narrative is incomplete and that the introduction of new technology is not a unique phenomenon, but rather a continuation of existing trends, and that the focus should be on empowering workers and addressing the underlying issues rather than just blaming AI for job displacement.

The great AI delusion is falling apart

The hype surrounding AI's ability to increase productivity is not matching real-world results, with some studies showing that AI tools can actually slow down tasks and generate more work. The true measure of AI's productivity benefits should consider the entire process, including verification and editing, rather than just individual tasks.

Research

Safer AI Agents Through Understanding and Evaluating Mobile UI Operation Impacts

Researchers are exploring the creation of autonomous agents that can manage daily tasks by operating user interfaces, and are investigating the real-world impacts and consequences of mobile UI actions taken by AI agents. A study was conducted to develop a taxonomy of these impacts and evaluate the ability of large language models to understand them, revealing both enhanced reasoning capabilities and significant gaps in classifying nuanced or complex categories of impact.

TaxCalcBench: Can AI file your taxes? (not yet)

Current AI models are unable to accurately file personal income taxes due to the complexity of the task, which requires understanding and applying vast amounts of text-based information. State-of-the-art models can only successfully calculate less than a third of federal income tax returns, even with simplified samples, highlighting the need for additional infrastructure to improve their performance.

Frugal Machine Learning for Energy-Efficient, and Resource-Aware AI

Frugal Machine Learning (FML) involves designing efficient and cost-effective ML models that minimize resource consumption, including computational resources, time, energy, and data. The field of FML explores strategies such as model compression, energy-efficient hardware, and data-efficient learning techniques to achieve acceptable performance while reducing resource usage, particularly in resource-constrained environments like edge computing and IoT devices.

Transformers without normalization

Transformers without normalization layers can achieve similar or better performance using Dynamic Tanh (DyT), a simple element-wise operation that replaces normalization layers. By incorporating DyT, Transformers can match or exceed the performance of their normalized counterparts across various settings, challenging the conventional understanding that normalization layers are essential in modern neural networks.

Gemini 2.5 Pro Capable of Winning Gold at IMO 2025 with Prompting

The International Mathematical Olympiad (IMO) poses uniquely challenging problems that Large Language Models (LLMs) struggle with, but using Google's Gemini 2.5 Pro with a self-verification pipeline and careful prompt design, 5 out of 6 IMO 2025 problems were solved correctly. This result highlights the potential of LLMs for complex reasoning tasks and the importance of developing optimal strategies to harness their full potential.

Code

Show HN: LLMs suck at writing integration code… for now

Superglue is a platform that enables users to integrate and orchestrate APIs using natural language, allowing agents to build deterministic workflows across various apps, databases, and APIs. It features automated schema mapping, drift detection, retries, and remappings, and provides a range of tools and features, including a lightweight proxy, LLM-assisted mapping, and self-healing schema drift, to make API workflows more reliable and efficient.

Show HN: Local Email Client for AI Horseless Carriages

DispatchMail is an open-source, AI-powered email assistant that helps manage your inbox by monitoring emails, processing them with an AI agent, and providing a web interface for managing drafts and responses. The application is currently in alpha stage, runs locally, and uses OpenAI for queries, with plans to improve and expand based on user feedback and interest.

LiteLLM Python SDK Proxy Server LLM Gateway Call 100 LLM APIs in OpenAI Format

LiteLLM is a platform that allows users to call various LLM (Large Language Model) APIs, including OpenAI, Huggingface, and VertexAI, using a unified format. It provides features such as retry/fallback logic, routing, and budgeting, and supports multiple providers, including OpenAI, Huggingface, and VertexAI, with consistent output and streaming capabilities.

Show HN: A tiny (480 LOC) AI coding assistant for your shell

TinyCoder is a command-line AI coding assistant that works within your existing shell environment, allowing you to send messages to an AI assistant and receive responses that can execute commands or ask follow-up questions. It supports various AI model providers, including Ollama, Google Gemini, and OpenRouter, and can be configured using environment variables to customize the AI model and provider.

Structllm – structured output support to any LLM provider

StructLLM is a Python library that provides structured output functionality for large language models (LLMs) from various providers, ensuring responses conform to a provided JSON schema using Pydantic models. It supports over 100 LLM providers through LiteLLM and allows for customization of parameters such as temperature and top_p settings to ensure consistent outputs.

    OpenAI prepares to launch GPT-5 in August, researchers find that Transformers can achieve similar performance without normalization layers, and Superglue enables users to integrate and orchestrate APIs using natural language.